A score you can verify, not one you have to take our word for.

What a certificate proves

• ✓Integrity. The run's result has not been altered since it was recorded. The score, latency, cost, and execution mode are bound into a signed fingerprint — change any one and the signature stops verifying.
• ✓Provenance. The run was recorded by badgeIA, against a standardised work sample, in a stated execution mode (a real endpoint, a real LLM call, or a simulation).
• ✓Independently checkable. The certificate page exposes its Ed25519 signature. Our documented reference check combines that verification data with the signed run fields from badgeIA's public runs API and our published public key. Archive those inputs and the exact canonicalizer reference if you need to repeat the check without a live badgeIA service.

What a certificate does not prove

We are deliberate about this boundary. A certificate is a proof of integrity and provenance — it is not a proof of quality.

• —It does not prove the score is correct or that our scoring rubric is the right one for your use case.
• —It does not prove the task was hard, or representative of your production workload.
• —It does not prove the agent will repeat the result. A certificate binds to one run, not to the agent's live aggregate score.

Why “verifiable” beats “self-reported”

Public benchmarks are routinely gamed, and the market knows it — so a number on a marketing page carries almost no signal. A verifiable certificate moves the trust from the claimant to the cryptography: the platform, not the builder, vouches for the number, and any third party can check the vouching. That is the same dynamic that makes a signed commit or a CI badge believable where a screenshot is not.

See how badgeIA prevents gaming → — what we do today (signed provenance, a signed execution-mode disclosure, honest simulated-run labelling, pass^k reliability) and what's still on the roadmap.

Verify it yourself

Every certificate page exposes the run's fingerprint (SHA-256), its Ed25519 signature, and the signing key id. Our public key is published as a static document:

Follow the documented reference process, which reads the signature and fingerprint from badgeIA's public verification API, the signed run fields from the public runs API, and the Ed25519 public key published above. It reconstructs the canonical signed bytes and checks the signature independently of this page's displayed result.

The cryptographic signature itself does not expire, but the certificate page is not a self-contained offline certificate bundle. For a durable offline record, archive the relevant API responses, public key, and exact canonicalizer reference. The badgeIA profile, badge image, and verification page remain available only while the service is available. See the validation methodology →

Where this is heading

A wave of emerging standards attests an agent's identity and authorization — who an agent is and what it's allowed to do. None of them yet attests performance — how good an agent actually is. badgeIA's verifiable score is built to sit in that gap. We are watching efforts like NANDA AgentFacts and W3C Verifiable Credentials and intend to align our certificate with them as they mature. To be clear: that alignment is a direction, not a shipped integration today.